Database Reference
In-Depth Information
fit the current clustering, a request for increasing the radius is sent to
a coordinator. The coordinator replies with the biggest radius it has
received from all other nodes. The local nodes maintain their heap such
that the effect of the most
l
dense clusters which appeared in history
solutions is kept, but also such that space is left for establishing new
clusters if there is a new trend in the input stream. The coordinator re-
ceives the local solutions
C
i
,radii
R
i
and radius increase requests from
the nodes. It continuously performs the Furthest Points algorithm on
the solutions
C
i
and keeps the largest radius received from all nodes.
The base station (server side) rotates the coordinator according to an
estimate of the residual energy in each node. EDISKCO determines a
(4 +
ε
)-approximation of the optimal global clustering. Empirically, it
was shown that the algorithm outperforms the centralized Global Paral-
lel Guessing algorithm that was proposed by Cormode
et al.
[16], with
regard to accuracy as well as energy consumption.
Incorporating energy saving techniques from sensor node clustering
into methods for distributed data analysis, like regularly switching the
role of the central coordinator, seem to be a fruitful area of future re-
search. This not only concerns the clustering of sensor measurements,
but also methods for classification and prediction, like the ones presented
in the following section.
3. Classification in Wireless Sensor Networks
Collaborative target classification is an active area of research in the
WSN community. US government funded projects through DARPA and
Department of Defense (DoD) are interested in a variety of sensor net-
works applications for modern warfare. One such classic application is
multi-vehicle tracking and classification using distributed wireless sen-
sor networks. The goal here is two fold. Since sensors are deployed
across the hostile terrain, the first goal is to develop collaborative mod-
els which use the data of all sensors and then deploy distributed data
mining techniques to build such models using low power consumption
and communication overhead. A major advantage of using such col-
laborative techniques is to bolster the inference of one node using the
posterior of the other node. In essence, if one node can validate a hy-
pothesis, then in makes more sense to use it for subsequent inferencing
rather than starting from scratch for each node. This forms the sec-
ond goal of such inferencing technique. Such a collaborative system was
developed and deployed by Meesookho
et al.
[40] for identifying and
classifying vehicle types from a convoy of vehicles. The paper shows
that using confidence boosting, which uses the posterior of one node to
